The Inception of the Virtual Autonomous Learner (VAL)

Inception of VALThe inception of VAL was conceived in an advanced reading class where I studied the current state of artificial intelligence (AI). The background research involved in this project was heavily interdisciplinary. Some of the major research areas were evolution, evolutionary psychology, experimental psychology, neurology, ecology, learning theories, educational technologies, computation, simulation, computational neural modeling, early child development, cognition, linguistics, computer science, and philosophies like Theory of Mind. Much of my research led to experimental psychology and the field of computational neural modeling and cognitive architectures.

In the beginning, I was overwhelmed and astonished with the variety of cognitive architectures that were currently being developed within reputable institutions1.

Even though, in the past decade, new technologies and research techniques have given cognitive sciences a more vivid picture of what is going on cognitively in our brains it is surprising how many perspectives there are with regards to how we think.

In another advanced reading class, where VAL’s theoretical groundwork was laid out, I was more focused most likely due to my growing biases in the field of cognitive architectures. In terms of different domains of architectures, I started to favour the connectionist approach over symbolic processing.

However, even after narrowing the focus of my research it became apparent that what was I was looking for did not exist. I was looking for the perfect cognitive architecture.

Slowly I realized having different perspectives on how to solve a problem is not necessarily a bad thing – as Marvin Minsky puts it in The Society of Mind, “The power of intelligence stems from our vast diversity, not from any single, perfect principle” (Minsky, 1987, p. 308).
Even though I was very impressed with the many simulations and their results concerning behaviour data and how comparable the data is to human data it is clear this field is still in its infancy stage.

After reviewing a dozen or so architectures, all different and all showing much research potential, it was obvious that there was a need for a tool that could easily experiment with different architectures.

Basically, the visceral nature of cognition led me to believe that the foundation of VAL’s construct should be built on a framework of a formative evaluation of the architecture’s development.

In 1990, Allen Newell called for a unified theory of cognition (Newell, 1990). Twenty years later there are more cognitive architectures than ever.

Perhaps what we need is a unified construct on which to learn about cognition?
In addition to creating a construct that can implement different architectures it was evident there was a need for an embodied agent. Therefore, my goal is create an embodied ecological computational cognitive construct (EECCC).

In an attempt to create a better embodied agent, a survey of multi-agent environments was conducted. The research revealed some common traits and ideological approaches that seemed logical. In terms of developing the construct, a set of principles emerged as the balance between biological realism and practicality. These principles cover three levels of analysis, (1) the body, (2) the mind, and (3) the environment. A by-product of this type of analysis is that researchers can take both the “top-down” and “bottoms-up” approach to creating a better agent.

The trilateral relationship of mind / body /environment is at the heart of this ecological perspective. Explicitly, I believe the fundamentals of knowledge is acquired by interactions with the environment.

The VAL construct is a nature vs. nurture playground. VAL’s motor control, sensory, and cognitive systems consist of individual units, which simulate the function of a biological neuron. These units are interconnected and form a network of neurons. How they are connected forms the topology of the network – this is the architecture.

The architecture is heavily biased and these biases create innate traits that will dictate how VAL will interact with its environment. High-jacking the popular aphorism “blank slate” (the notion that the mind has no innate traits), I am using it to describe the initial state of the agent’s architecture – before learning occurs.

I soon realized, in addition to the ecological perspective, evolution offers us a rich insight on both the physical and cognitive development of human behaviour. Many researchers are using evolutionary psychology as a way to reverse-engineer mental constructs (Tooby & Cosmides, 2005; Pinker, 1997).

I researched different stages of vertebrate evolution to inquire about the environmental demands of the organism at that particular point in evolution, asking what were the physical and cognitive demands the organism needed to survive, then using that information to model the initial state of the neural network’s architecture. Therefore, the blank slate that VAL uses to build a mental model of the world is not only influenced by ecology, but also greatly influenced by evolutionary processes.

  1. See A Cognitive Architecture Primer for a short list

Tags: , , ,

Leave a Reply